Exbrain 2.0 — A Self-Growing External AI Brain

An AI knowledge system that remembers, compiles, verifies, and evolves on its own.
Claude Code × Obsidian × a 4-layer architecture

🇯🇵 日本語 · Inspired by Karpathy's LLM Wiki

What is Exbrain?

There's exactly one reason the most powerful AI model still gives you mediocre output: **it knows nothing about you.**

Exbrain builds an "external brain" that Claude Code reads from and writes to, on top of Obsidian (which is really just a folder of Markdown files). Your days, clients, decisions, and lessons accumulate and get compiled automatically, so that **every session starts already knowing where you are right now.**

• Your memory lives on your own disk, as plain text. No vendor lock-in.
• It keeps working with your laptop closed (via cloud-side scheduled tasks), and you can read it on your iPhone (via iCloud sync).
• Whatever model you swap in next, the brain survives.

What Changed in v2

The single biggest lesson from three months of running v1 (SOUL/MEMORY/DREAMS + clips):

**"Raw logs pile up on their own, but compiled knowledge rots on its own."**

Daily notes and clips (the raw layer) grew fatter every day through automation, while client pages and indexes (the wiki layer) went stale within weeks. v2 fixes this with a **nightly compiler.**

	v1	v2
Structure	SOUL/MEMORY/DREAMS + a set of folders	4-layer model (raw → wiki → digest → identity)
Front door	None (just CLAUDE.md)	INDEX.md (map + freshness dashboard)
Raw → knowledge compilation	Manual (stalls quickly)	Nightly compile runs every night (cheap model)
Quality control	weekly-sync.sh (never actually ran)	Weekly lint (regist